Data-driven ensemble forecasting with the AIFS
Alexe, Mihai ; Lang, Simon ; Clare, Mariana ; Leutbecher, Martin ; Roberts, Christopher ; Magnusson, Linus ; Chantry, Matthew ; Adewoyin, Rilwan ; Prieto-Nemesio, Ana ; Dramsch, Jesper ; Pinault, Florian ; Raoult, Baudouin
Data-driven weather forecast models are a promising addition to physics-based numerical weather prediction (NWP) models. ECMWF now runs the Artificial Intelligence Forecasting System (AIFS) in an experimental real-time mode. It is run four times daily and is open to the public under ECMWF's open data policy. This AIFS version (henceforth referred to as 'deterministic AIFS') is trained to produce forecasts that minimise mean squared error (MSE) up to 72 h into the forecast. The MSE optimisation leads to excessive smoothing and reduced forecast activity (Lang et al., 2024(a)). This is detrimental to ensemble forecasts, which rely on a realistic representation of the intrinsic variability of the atmosphere.<br />In this article, we describe two training approaches for data-driven forecast models to produce skilful ensemble forecasts: <em>diffusion training</em> (Karras et al., 2022, and Price et al., 2024), where the forecast is the result of a denoising task, and <em>probabilistic training with a proper score objective adjusted for the finite ensemble size</em>, such as the fair continuous ranked probability score (fair CRPS; Leutbecher, 2019, and Kochkov et al., 2024).</p>
Accès à la notice sur le site du portail documentaire de Météo-France